The Estimation of "Transfer Functions" of Quadratic Systems

Author(s): Leo J. Tick

Source: Technometrics , Nov., 1961, Vol. 3, No. 4 (Nov., 1961), pp. 563-567
Published by: Taylor & Francis, Ltd. on behalf of American Statistical Association and
American Society for Quality

Stable URL: https://www.jstor.org/stable/1266562

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Taylor & Francis, Ltd. , American Statistical Association and are collaborating with JSTOR to
digitize, preserve and extend access to Technometrics

This content downloaded from

143.107.3.15 on Tue, 08 Jun 2021 23:57:30 UTC
All use subject to https://about.jstor.org/terms
VOL. 3, No. 4 TECHNOMETRICS NOVEMBER, 1961

The Estimation of "Transfer Functions" of

Quadratic Systems'
LEO J. TICK
Engineering Research Division, New York University

1. SUMMARY AND INTRODUCTION

In recent years the use of stationary random inputs as a forcing function to

experimentally determine the transfer function of linear systems has become
widespread. The procedure involves the measurement of spectra and cross-
spectra between the input and output and the formation of the proper ratio.
There are two basic reasons for using random testing functions. For many situa-
tions, particularly mechanical ones, these inputs are easier to generate than say
steps or impulses and frequently they can be made to more closely approximate
the in-service input. (The latter attribute is an advantage since the linearity of the
system may not be complete but may be sufficiently so over the operating range
of interest). A simple measure of linearity is immediately available i.e., the
coherency.
Although many experimenters try to generate a Gaussian stationary process
for the forcing function this is not necessary from an expected value veiwpoint.
Actually the probability structure plays no role in the general logic of the de-
tailed procedure since only second moment characteristics of the process are
relevant to the expected value calculations. However, a Gaussian input can be
convenient since sometimes the evaluation of the variability of the estimates
is simplified. In studying higher order systems by driving them with a random
forcing function this use of a Gaussian process makes the calculations of the
expected values considerably easier. In this paper we shall extend the spectral
techniques of transfer function estimation of linear systems to time invariant
quadratic systems when a stationary Gaussian process is used as a driving
function.

2. TIME INVARIANT QUADRATIC SYSTEM

If X(t) is the input to a system S and Y(t) is its output, S will be said to be a
time invariant quadratic system if

(1) Y(t) = l(t - r)X(T) dr + q(t - rT , t - r2)X(Tr)X(r2) dr, dr2

The functions I and q are to be viewed in a general since in that they may con-
tain delta functions and derivatives thereof. It is possible to avoid considera-
tion of such functions by giving definitions in the Fourier domain and we do that
1 This research supported by David Taylor Model Basin under Contract Nonr 285(17). Re-
production in whole or in part is permitted for any purpose of the United States Government.
563

This content downloaded from

143.107.3.15 on Tue, 08 Jun 2021 23:57:30 UTC
All use subject to https://about.jstor.org/terms
564 LEO J. TICK

below. We refer to both I and q jointly as the impulse response function of S.

Without loss of generality we can assume q(a, b) = q(b, a). The quantities L
and Q, if the integrals exist, are defined by

(2) (r) = fe- 'L(w) dw,

(3) q(r, , r2) =- f e- i ( T1+ 22)Q(Wl , C2) dcO

We refer to them as the transfer function of the system. L a
the transfer and impulse functions of a linear system. Let X(t) a
representation

(4) X(t) = f ett d().

Such a representation is possible if X(t) is a statio
an integrable deterministic function. Equation (4)

(5) y(t) = f eitL(o) d9(w) + ff eit(w1+2)Q(W1 , 2) d(W,) d 2(w,).

It is possible, and easier from a standpoint of rigor to use (5) as the definition
of a time invariant quadratic system.
Before we consider the estimation of L and Q it is well to say a few words
about how the form can be expected to arise. It can arise by assumption be-
cause this is the nature of the system. It also can arise because the extent of
non linearity in the system is sufficiently small so that a perturbation expansion
may be truncated after the quadratic terms. In a way (1) or (5) represents a
kind of truncated polynomial expansion. Weiner [1] uses this type of expansion
(5) to represent an arbitrary random process from Brownian motion process.
He refers to the terms in the expansion as polynomial functionals.

3. ESTIMATION OF THE TRANSFER FUNCTION

Suppose that X(t) is a stationary Gaussian process with zero mean and spectral
density s,,(w) and covariance function Rfi(r). (It is sufficient to assume that
the third and fourth cumulant functions of X(t) are identically zero but for
simplicity we make the more restrictive assumption.) Then

E[Y(t) = ff q(t - Tr t - T2)RI(7r,2- 7) dT1 d2r

(6)

= J Q(c, -cO)s() dw = M,

The cross-spectrum between the input and output s,,(c) is easily computed to
be, if s,,(o) is the input process spectrum,

(7) sZ(c) = s(o)L().

Therefore

(8) L(wc) = sxv(w)/s (C)

This content downloaded from

143.107.3.15 on Tue, 08 Jun 2021 23:57:30 UTC
All use subject to https://about.jstor.org/terms
 "TRANSFER FUNCTIONS" OF QUADRATIC SYSTEMS 565

The usual procedure for estimating the transfer function of a linear system w
calculate the linear component of the transfer of a quadratic system. The thi
moment function about the mean of the input, input, and output, is

E[X(t + t2)X(t + tl) Y(t) - E[Y(t)]}] = M3(tl , t2)

= q(t - r, t - r2)E[X(rl))XrX(t + tl)X(t + t2)

(9) - R )(0)R(7r2 - T7)]}

= / q(t - 71, t - 72)[R2(t + tl - rT

+ R,x(t + tl - T2)R,(t + t2- -r)]

= 2 Jf ei (t"1 + t 2WQ( ,( 2)s1).(WI)s(W2) dwl dw2

In the above calculation we used the property that the third and fourth cumulant
functions are zero. Following J. W. Tukey we call the Fourier Transform of
M(t , t2) the cross bi-spectrum. Therefore

(10) Q(l , 2) = Cross Bi-Spec. (cw , w2)

2s,x((1)sx(C2)
where, of course,

(11) Cross Bi-Spec (w, , 2) = (I2 fJ e-i"'I+ 2'2)lM(tl , t2) dt, dt2

The cross bi-spectrum defined above has certain obvious symmetry properties.
Some are C.B.S. (l , Co2) = C.B.S. (w2 , cw), since M3(t2, t,) = M(t, ,t2). Also
E{X(t)X(t + tl)[Y(t + t2) - EY(t + t2)]} = M3(t2 - tl, - t2). Equations
(8) and (10) provide a procedure for estimating the transfer function of a time
invariant quadratic system when a stationary Gaussian process is used as a
forcing function. Now

(12) Cov [Y(t)Y(t + r)] = f e" L L(w) 2 sx(w) dw

+ 2 ff eiT(+' I Q(wl , w ) 12 Sxx(Wl)sxx(w2) dwl d2.

The Fourier transform of (12) is the spectral density, and is

I L(w) 12 s..(W) + 2 Jf Q(co - , 12 x(c - r)sx(r) dr.

4. A MEASURE OF QUADRATIC COHERNECE

If it may be assumed that there is virtually no noise in the measuring system,

and if q = 0 we have a simple measure of the degree of linearity of the system.
This quantity is the coherence defined by

(13) coh2 (c) = [co-spec]2 + [quad-spec]2

(specl) (spec2)

This content downloaded from

143.107.3.15 on Tue, 08 Jun 2021 23:57:30 UTC
All use subject to https://about.jstor.org/terms
566 LEO J. TICK

The coherence will be one at any frequency at which there is a linear relationship
and less than one otherwise. It would be convenient to have a measure (with
the reader's indulgence) of the "Quadracity" of the system i.e., a measure of
"goodness-of-fit" of the quadratic model. M. A. Woodbury has observed that
the definition of linear coherency is the ratio of the output spectrum determined
from cross-spectral techniques assuming the existence of a transfer function of
the output spectrum. This observation leads immediately to a measure of
"quadratic coherency." It is

(1 quad. coh (c) =- 1 {s +( 1 I C.B.S. ( - X) dX

In the above we have used (3, 8, 10, 11).

It can easily be shown that (14) is less than or equal to one. The proof is
exactly the same as the proof that the linear coherency is less than or equal to
one. Eq. (14) will be one when the system is quadratic.

5. CONCLUDING REMARKS

It is possible to take a somewhat different and perhaps more general view-

point on the estimation equations (8, 10). If the quadratic model (5) represents
the system then these equations are a means of estimating the parameters of
the system. However, for an arbitrary system they are also a means of "fitting" a
quadratic model of the form (5) to this arbitrary system* where the quad-
coherency, (14), provides a measure of the success of the fitting. Therefore,
the procedures contained herein can be used to obtain a "general" steady state
solution, within the quadratic class (5) of, say, a non-linear differential equation
be experimental means. This can be done by setting up the equation on, say,
an analog computer and driving it with Gaussian noise. What fitting criterion
this procedure conforms to is an open question.
The reader should notice that two problem areas related to the use of these
estimation procedures have been ignored. They are: computation procedures,
bias and variability. There are two general procedures for making these calcula-
tions. The first involves the calculation of the sample third moment, smoothing
and taking Fourier transforms. It will be observed upon a little analysis that
in contradistinction, to the case of spectral estimation, most of the effort in
bispectral computation is in the Fourier transform. Therefore, it may be reason-
able in some problems not to compute the transform over the entire range.
The second procedure is analogous to computing a spectrum by using a narrow
band filter and squaring and averaging the result. The filter ising thee n tuned through
the frequency bands of interest. For bispectral estimation the series are filtered
through three narrow-band filters and the triple products taken and smoothed.
Here one obtains only the bispectrum at the points of interest. The reader
should bear in mind that bispectral computations are lengthy and should only
be attempted on the fastest equipment. As for variability, we have e performed
* This is similar to approximating an arbitrary curve by a polynomial of a given order.

This content downloaded from

143.107.3.15 on Tue, 08 Jun 2021 23:57:30 UTC
All use subject to https://about.jstor.org/terms
 "TRANSFER FUNCTIONS" OF QUADRATIC SYSTEMS 567

some computation on synthetic series and the indications are that

five times (Tukey suggests ten) as much data as in the spectr
hoped to make a more detailed report in a later paper.

ACKNOWLEDGMENTS

The author wishes to thank R. F. Drenick and J. Nadler of The Bell Tele-
phone Laboratories for their criticisms of an early manuscript, and to the
referees for pointing out some errors.
REFERENCES

[1] WEINER N. (1958), Non-Linear Problems in Random Theory, Technology Press.

This content downloaded from

143.107.3.15 on Tue, 08 Jun 2021 23:57:30 UTC
All use subject to https://about.jstor.org/terms